[Insight-users] Fuzzy K-means clustering

[Richard Beare]

Hi,
This is an alternative approach that I've found useful in the past. If
thresholds are close to doing what you are after then it has a good
chance for working for you.

Since you have only two classes in your image, presumably one is light
and the other dark. It should be relatively easy to estimate a pair of
"conservative" thresholds - one of which correctly selects the really
bright parts and the other which selects the really dark parts. If the
bright regions are small blobs, the resulting segmentation should
appear undersized at this point. There is also likely to be a large
area of unassigned voxels that are of middle brightness - the next
step is to assign them. There are two standard approaches I usually
use at this point - either watershed transform or seeded region
growing (the greedy Adams and Bischof version). In many cases the
results will be similar. The process is to generate a marker image
from your two conservative segmentations - e.g make the output of the
dark threshold 1 and the output of the bright threshold 2 and combine
the pair into a single image (by using a max or add operation). This
will be the marker image. Feed this into the morphological watershed
with the gradient of the raw image as the control image, or into
seeded region growing with the raw image as the control. Both
processes will assign classes to all unassigned voxels in what is
often a very reliable way.

What I've just described is a classical morphological segmentation
procedure as described by Meyer. The main variants are on how you
generate markers.

I hope this helps

On Thu, Jul 24, 2008 at 5:30 PM, Anja Ende <anja.ende at googlemail.com> wrote:
> Hello everyone,
>
> I want to use the K-means classifier to do some segmentation. I want
> to classify my image into just 2 different classes (fluid and fat). I
> already have a fluid suppressed MR images and the fat voxels are
> supposed to be bright. I want to use an automatic segmentation method
> to do the job.
>
> I tried the Otsu segmentation which gives decent results some of the
> time but I would like to improve it and was thinking of trying other
> algorithms to see if I can improve it.
>
> My questions are:
> - Is the k-means classifier in ITK an implementation of the fuzzy
> k-means clustering? If no, is there an implementation of the fuzzy
> k-means classifier in ITK.
> - If I use the k-means classifier with the idea of only dividing the
> image up into 2 classes... Is it the same as doing an Otsu
> segmentation?
>
> I have a pretty urgent requirement regarding this and any help from
> some segmentation experts would be greatly appreciated.
>
> Cheers,
> Anja
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>